Service Life Prediction of Organic Coatings - ACS Publications

container. The liquid in the container caused pitting corrosion in the crevice area. (indicated by the arrow), and the pit grew until it perforated th...
1 downloads 0 Views 1MB Size
Chapter 12

Downloaded by UNIV OF CALIFORNIA SANTA BARBARA on May 29, 2018 | https://pubs.acs.org Publication Date: April 15, 1999 | doi: 10.1021/bk-1999-0722.ch012

Reliability Engineering: The Commonality Between Airplanes, Light Bulbs, and Coated Steel William Stephen Tait S. C. Johnson and Son, Inc., 1525 Howe Street, Racine, WI 53403-5011

Failure comes in a variety of different forms, depending on the system and the environment to which it is exposed. However, failure has attributes that are common to all forms of failure. The attributes of failure are: 1) chemical and physical phenomena initiate small flaws in materials, 2) flaws grow under certain environmental conditions, 3) flaw growth rates can be used to estimate time of failure, 4) failure occurs when flaws grow to a critical size, or damage accumulates at a high enough level to cause failure, or the flaw can be observed by the unaided eye and considered a failure, 5) the number of observed failures in a population is not a continuous function of time, 6) the cumulative number of failures in a population is a continuous function of time, 7) the magnitude of cumulative failure in a population is influenced by the environment, and 8) cumulative failure data can be transformed into a linear form and the data extrapolated outside of its range (within reason) to estimate population failure level. These attributes can be used to estimate long-term failure times and associated levels of failure using reliability statistics. Failure levels for aerosol containers can be estimated to within an order of magnitude of actual failure levels. Every industry strives to make products that will perform for a designated service lifetime without the level of failure exceeding a specified maximum. A specified maximum level of failure is often referred to as an acceptable quality level, or AQL. * An example of an AQL and its associated service lifetime is 1% failures prior to five years. One method for estimating failure times and associated levels is a long-term exposure test. Samples are exposed to service conditions; periodically inspected for failures; and the number of failures at each inspection time are recorded.

186

© 1999 A m e r i c a n C h e m i c a l Society

Bauer and Martin; Service Life Prediction of Organic Coatings ACS Symposium Series; American Chemical Society: Washington, DC, 1999.

Downloaded by UNIV OF CALIFORNIA SANTA BARBARA on May 29, 2018 | https://pubs.acs.org Publication Date: April 15, 1999 | doi: 10.1021/bk-1999-0722.ch012

187 Long-term exposure tests can take several years to complete because they are conducted until all samples have failed. The objective of an exposure test is to determine if the desired service lifetime is achievable without exceeding acceptable failure levels. Unfortunately global competitive pressures usually prevent conducting tests that take several years to complete, and results from a censored test must be used to estimate long-term failure levels, or verify that service lifetime will be achieved without exceeding acceptable failure levels. A censored test is one in which samples are removed from test prior to failure, or the test is terminated prior to failure of all samples.^ Different Types of Failure There are many different types of failure. Figure 1 contains an example of a cracked empennage cap on a single engine airplane (a failure in progress). Mechanical force is applied to the empennage when an airplane begins to fly, and the force is removed when the airplane lands. The mechanical force applied during each flight-cycle initiates fatigue cracks like that shown in Figure 1, and causes the cracks to grow until there is enough damage to cause failure. Cyclic mechanical stress is analogous to cyclic stresses that result from daily and climatic changes in weather.

Figure 1. Cracked empennage cap on a single engine airplane Figure 2 contains a photograph of a burned-out light bulb. A thermal stress is applied to the bulb filament when the bulb is turned-on, and the thermal stress is removed when the bulb is turned-off. Heat produced by electrical resistance also causes loss of filament material during bulb use. The cyclic thermal stress and thermal corrosion cause the bulb filament to crack, and the crack grows during each use cycle until the filament fails (breaks). Thermal cycling also occurs during weathering of coated metals.

Bauer and Martin; Service Life Prediction of Organic Coatings ACS Symposium Series; American Chemical Society: Washington, DC, 1999.

Downloaded by UNIV OF CALIFORNIA SANTA BARBARA on May 29, 2018 | https://pubs.acs.org Publication Date: April 15, 1999 | doi: 10.1021/bk-1999-0722.ch012

188

Figure 2. Burned-out light bulb. Figure 3 contains a photograph of a corroded automobile door panel. Water absorbs into the paint and diffuses to the paint-metal interface to cause blistering. Blistering causes separation of the paint from the door panel, and more water diffuses through the paint to accumulate inside the blister. Water accumulation in a blister causes metallic corrosion that progresses under the paint, enhancing both growth of the blister and separation of paint from the metal substrate.

Figure 3. Rusted automobile door panel.

Bauer and Martin; Service Life Prediction of Organic Coatings ACS Symposium Series; American Chemical Society: Washington, DC, 1999.

189

Downloaded by UNIV OF CALIFORNIA SANTA BARBARA on May 29, 2018 | https://pubs.acs.org Publication Date: April 15, 1999 | doi: 10.1021/bk-1999-0722.ch012

Figure 4 contains a photograph of the bottom portion of a perforated aerosol container. The liquid in the container caused pitting corrosion in the crevice area (indicated by the arrow), and the pit grew until it perforated the container metal.

Figure 4. Perforated aerosol container The different types of failure in Figures 1 through 4 have several attributes in common: 1. There are chemical and physical phenomena, such as electrochemical corrosion and thermal or mechanical stress, that initiate small flaws in materials 2. Flaws grow under certain environmental conditions 3. Failure occurs when flaws grow to a critical size and; a) damage accumulates at a high enough level to cause mechanical failure, or b) the flaw is considered to be a failure because it is large enough to be observed by the unaided eye 4. Flaw growth rates can be used to estimate time of failure 5. The number of observed failures in a population is not a continuous function of time 6. The cumulative number of failures in a population is a continuous function of time 7. The magnitude of cumulative failure in a population is influenced by the environment 8. Cumulative failure data can be transformed into a linear form, and the data extrapolated outside of its range (within reason) to estimate the time and level of failure for a population. These attributes of failure can be used with reliability statistics to predict service lifetime from short-term censored tests. Service lifetime can be mathematically expressed with the following general equation:

Bauer and Martin; Service Life Prediction of Organic Coatings ACS Symposium Series; American Chemical Society: Washington, DC, 1999.

190 EFTj =

Ti + _S*_

(1)

Downloaded by UNIV OF CALIFORNIA SANTA BARBARA on May 29, 2018 | https://pubs.acs.org Publication Date: April 15, 1999 | doi: 10.1021/bk-1999-0722.ch012

rate Where the rate term in Equation (1) is flaw growth rate; EFTj is the expected failure time (service life) for an individual unit; T\ is the initiation time for the flaw that causes failure; and S* is the flaw size that causes or constitutes failure. Initiation time (Tj) is often a distribution of times, and defect growth rate ("rate" in Equation 1) is a function of the material and its environment. Linear growth rates have been observed for metal pitting corrosion such as that in Figure 4;3>4,5 exponential growth rates are expected for metal cracking like the examples in Figures 1 and 2;^ and linear7,8,9 exponential*° growth rates have been reported for organic coating blistering such as that in Figure 3. It will shown that an approximation of Equation (1) can be used to estimate failure time and cumulative failure level for aerosol container failure by pitting corrosion such as that shown in Figure 4. The correlation between predicted and real-time results for aerosol containers will be discussed at the end of this paper, along with some key elements that are needed for any successful reliability program. o r

Attributes of Failure A brief discussion of the attributes of failure will provide background for how they are used with reliability statistics to model failure levels. 1) There are Chemical and Physical Phenomena, such as Electrochemical Corrosion and Thermal or Mechanical Stress, that Initiate Small Flaws in Materials. Phenomena like electrochemical corrosion or stress can initiate flaws such as pits, paint blisters, and cracks in materials. Flaws typically initiate in microscopic areas. 11,12,13 Consequently the onset of failure is often difficult to detect with the unaided eye, and it takes time before the flaw size can be observed with the unaided eye, or measuring equipment must be used to detect flaws in materials. 2) Flaws Grow under Certain Environmental Conditions. It is well known that flaws like fatigue cracks, paint blisters, and corrosion pits all continue to grow as long as the environment that initiated the flaw, or conditions that enhance flaw growth, are not removed from the material.^ 14,15 3) Failure Occurs when Flaws Grow to a Critical Size; or Damage Accumulates at a High Enough Level to Cause Mechanical Failure; or the Flaw is Considered to be a Failure because it is Large Enough to be Observed by the Unaided Eye. Failure is defined by the type of service. For example, a steel bolt will break from tensile stresses when a pit or stress crack reduces bolt cross section to where there is no longer enough metal to support the applied load. Blisters on painted metals are considered failure when blister size is large enough to be observed with the unaided

Bauer and Martin; Service Life Prediction of Organic Coatings ACS Symposium Series; American Chemical Society: Washington, DC, 1999.

191 eye. Both of these examples encompass aflawthat has reached a size at which failure occurs.

Downloaded by UNIV OF CALIFORNIA SANTA BARBARA on May 29, 2018 | https://pubs.acs.org Publication Date: April 15, 1999 | doi: 10.1021/bk-1999-0722.ch012

4) Flaw Growth Rates can be used to Estimate Time of Failure. Flaws have growth rates that can be estimated or measured. Expected Failure Time (service lifetime) of an individual component is equal to the time that it takes for a defect to initiate in the material and grow to a critical size (S*) that causes failure, or is considered to be failure. 5) The Number of Observed Failures in a Population is not a Continuous Function of Time. The number of failures observed at different times in a population are seemingly random. Figure 5 contains a histogram for the number of aerosol container failures observed at different times during a long-term test. It can be seen in Figure 5 that the number of failures do not appear to follow a continuous trend, indeed, there are several times during the test when no failures were observed. 14

0

40

80

120

160

200

240

Exposure time (days)

Figure 5. The number of failures observed at a given time appear to be random 6) The Cumulative Number of Population Failures (or cumulative percent) is a Continuous Function of Time. Data like that in Figure 5 can not be extrapolated outside of test times. Fortunately the cumulative number of failures is a continuous function of time, as demonstrated in Figure 6, where data from Figure 5 are plotted in the cumulative form. Numbers of failures were normalized by converting them to percent of total test population. This conversion was done so that results from test parameters having different numbers of samples could be compared to each other. The cumulative form of failure in Figure 6 can be extrapolated to times outside test times because failure level is a continuous function of time.

Bauer and Martin; Service Life Prediction of Organic Coatings ACS Symposium Series; American Chemical Society: Washington, DC, 1999.

Downloaded by UNIV OF CALIFORNIA SANTA BARBARA on May 29, 2018 | https://pubs.acs.org Publication Date: April 15, 1999 | doi: 10.1021/bk-1999-0722.ch012

192

Ca* solution

1

2

t i t

******

ft + +



+

Deionized water 25

50

75

100

125

150

Days of Exposure Figure 7. Changing the environment changes failure times 7) The Magnitude of Cumulative Failure in a Population is Influenced by the Environment. The magnitude of cumulative failure is determined by the growth rate of the phenomena that causes failure, and this growth rate is determined by the material and the environment to which the material is exposed. 16 Figure 7 contains several failure curves for metal containers exposed to various salt solutions having 0.1 molar concentrations of salt in deionized water. ^ It can be seen that the failure

Bauer and Martin; Service Life Prediction of Organic Coatings ACS Symposium Series; American Chemical Society: Washington, DC, 1999.

193

Downloaded by UNIV OF CALIFORNIA SANTA BARBARA on May 29, 2018 | https://pubs.acs.org Publication Date: April 15, 1999 | doi: 10.1021/bk-1999-0722.ch012

level for the calcium solution was higher than those for potassium, sodium or lithium solutions. Water that was essentially ion-free (deionized water) had the lowest failure levels for any given time.

100

Expected failure time (days) Figure 8. An example of expected percent failures plotted as a function of associated expected failure times on Weibull probability paper 8) Cumulative Failure Data can be Transformed into a Linear Form. Data like that in Figures 6 and 7 can be transformed into a linear form with probability density functions (pdf) such as extreme value or Weibull functions. Figure 8 contains an example of failure data that have been transformed by plotting the data on Weibull probability paper. It can be seen in Figure 8 that the data are reasonably linear, and can be extrapolated within reason to test times that are outside of those for the data. For example, the data in Figure 8 are extrapolated to find the time at which 1% failures are expected to occur. There are other probability density functions that can be used to transform data into a linear form, 1? and it is advisable to use the pdf that best fits the data for an extrapolation like that illustrated in Figure 8. Case History: Using Failure Attributes to Model Failure of Aerosol Containers Aerosol containers are a common form of consumer packaging. Indeed, there were 2.8 billion metal aerosol containers produced in the United States in 1996.18 Failure for aerosol containers is defined in this paper as when a pit perforates the container to cause it to leak. Figure 4 has an example of an internally coated container that failed from pitting corrosion in the bottom double seam of the container. The algorithm used to model container failure time for pitted censored samples is contained in Figure 9. The pitting rate for a given container-formula system is estimated by dividing maximum pit depths for individual containers by the

Bauer and Martin; Service Life Prediction of Organic Coatings ACS Symposium Series; American Chemical Society: Washington, DC, 1999.

194

Downloaded by UNIV OF CALIFORNIA SANTA BARBARA on May 29, 2018 | https://pubs.acs.org Publication Date: April 15, 1999 | doi: 10.1021/bk-1999-0722.ch012

Maximum sample Pit Depth

Sample pitting rate = (Depth)/(Exposure Time)

Select highest rate among aU test samples

Remaining metal thickness = (sample thickness) - (pit depth)

Sample time before failure = (remaining metal)/(highest observed rate)

EFT = (Sample time before failure) + (Exposure time) Figure 9. Algorithm used to calculate expected failure time from pit depth and exposure time. corresponding exposure time, and selecting the highest rate among all samples as the best estimator for the system pitting rate. Pit depth is subtracted from metal thickness to get the metal remaining under the deepest pit. Remaining metal thickness is divided by the highest pitting rate to obtain the expected remaining time-beforefailure for each container. Expected failure time is calculated using a modification of Equation (1): EFTj = ET}

+

remaining metal

(2)

highest observed rate Where ET} is the exposure time for individual samples; remaining metal is the amount of metal left under a pit; and the highest observed rate is the highest pitting rate observed among all samples in the test. Pitting rate is estimated by dividing exposure time into the pit depth. Equation (2) is an approximation for Equation (1). This approximation is necessary because it is very difficult and time consuming to

Bauer and Martin; Service Life Prediction of Organic Coatings ACS Symposium Series; American Chemical Society: Washington, DC, 1999.

195 obtain the magnitude for Tj in Equation (1). It will be demonstrated that Equation

Downloaded by UNIV OF CALIFORNIA SANTA BARBARA on May 29, 2018 | https://pubs.acs.org Publication Date: April 15, 1999 | doi: 10.1021/bk-1999-0722.ch012

(2) is a reasonable approximation for Equation (1). The basic assumptions used to build the algorithm in Figure 9 are: 19 1. pits nucleate at different times 2. there is an intrinsic pitting rate for a given metal in a specific environment 3. the deepest pit observed among all samples examined at a given exposure time is the oldest pit, and thus the best estimator of the intrinsic pitting rate. Failure data were gathered from static storage tests in which aerosol containers were filled with the desired liquid, pressurized and stored in constant temperature rooms for one to several years 20 Static storage tests were censored because containers were periodically removed from constant temperature rooms; de-pressurized; emptied; open and inspected for corrosion. Constant temperature rooms were typically maintained at 21°C and a higher temperature such as 38°C. The algorithm used to model expected failure level for each expected failure time is contained in Figure 10. This algorithm follows the statistical procedure discussed by Kapur and Lamberson for graphical analysis of censored tests.21 Assign temporary EFTs to Censored Samples J Arrange all EFTs in ascending order

.

4

.

Calculate sample increments for determining rank order

4 Calculate median ranks for pitted samples J Calculate probabilities for pitted samples

Estimate population critical age | Figure 10. Algorithm for statistical treatment of censored test data. EFT is the expected failure time that is estimated from pit depth; growth rate; and container exposure time. Critical age is the time where the specified maximum failure level for a population is observed. Table 1 contains comparisons between expected failure levels (EFLs), and observed failure levels 20 it can be seen that the estimated EFL is typically within an order of magnitude of the observed failure level. The accuracy of expected failure times and their associated EFLs can be improved if the actual container pitting rate is known. Unfortunately the only way to

Bauer and Martin; Service Life Prediction of Organic Coatings ACS Symposium Series; American Chemical Society: Washington, DC, 1999.

196

Downloaded by UNIV OF CALIFORNIA SANTA BARBARA on May 29, 2018 | https://pubs.acs.org Publication Date: April 15, 1999 | doi: 10.1021/bk-1999-0722.ch012

estimate pitting rates for aerosol containers is to open containers and measure the pit depths, and assume the deepest pits were growing during the entire exposure time. The assumption about pit growth time causes inaccuracy in the estimated pitting rate, causing some of the differences between actual failure levels and EFL magnitudes observed in Table 1. The use of only the Weibull pdf to model the data also contributes to the differences between actual and EFL magnitudes. Evaluation of different probability density functions is a topic that is being pursued at this time and will be reported in a later publication. Table 1. Comparisons between estimated and observed failure levels EFL Observed failure levels 11.5% 5% 16.7 % 18.5% 17.7 % 33.3 % 9.2 % 22.2 % 8.1% 8.3 % 36.8 % 30% 55.6% 55.9 % 63.6% 64.4 % 5.6% 8.3 % 12.6 % 8.3 % 16.7 % 13.5 % 8.3 % 12.5 % 6.7 % 10% 41.7% 41.3% 23.8 % 25% 23% 14.3 % 29.8 % 33.3 % Successfully using Reliability Statistics There are a few key elements that should be incorporated into any reliability program: 1. Use replicate samples for each test parameter 2. flaw size that causes failure should be known or determined 3. flaw growth rate should be determined or estimated for each test parameter so that failure times can be estimated for censored samples 4. Use the pdf that best fits the data to estimate EFL Summary The attributes of failure can be used to obtain expected failure levels (EFLs) that are the same order of magnitude as actual failure levels. The algorithms described in this paper can be used for a variety of failure types where flaw growth rate is known or can be estimated, and the level of damage that causes failure is known.

Bauer and Martin; Service Life Prediction of Organic Coatings ACS Symposium Series; American Chemical Society: Washington, DC, 1999.

197

Downloaded by UNIV OF CALIFORNIA SANTA BARBARA on May 29, 2018 | https://pubs.acs.org Publication Date: April 15, 1999 | doi: 10.1021/bk-1999-0722.ch012

More accurate EFLs for aerosol container failure can be obtained if the actual pitting rate is measured without having to destroy containers to locate and measure pit depth. There are several probability density functions (pdf) that can be used to model failure data, so that data can be extrapolated (within reason) outside of the original data set. It is highly recommended that the best fit pdf data be used for extrapolation. Carte blanche use of the same pdf for all data can lead to inaccurate estimates for EFL. References 1. ANSI/ASQCZ1.4-1993; American Society for Quality Control, Milwaukee WI; 1993; p. 2 2. Nelson, W. Journal of Quality Technol; 1970; 2(3); p. 126. 3. J. E. Strutt, J. R. Nichols, and B. Barbier, Corrosion Science, 1985, 25(5), ppl. 305-315 4. J. W. Provan and E. S. Rodriguez III, Corrosion, 1989, 45(3), pp. 178-192 5. D. L. Crews, Galvanic and Pitting Corrosion - Field and Laboratory Studies, ASTM STP 576, American Society for Testing and Materials, Philadelphia PA, 1976 pp. 217-230 6. S. T. Rolfe and J. M. Barsom, Fracture and Fatigue Control in Structures, Prentice-Hall Inc., Englewood Cliffs NJ, 1977, p. 115 7. J. W. Martin, E. Embree, andW. Tsao, J. Coatings Tech., 1990, 62(790), pp. 2533 8. Leidheiser Jr., H. J. Adhesion Sci. Tech.; 1987; 1(1); pp. 79-98 9. Leidheiser Jr., H. and Wang, W., J. Coatings Tech., 1981, 53(672), pp. 77-84 10. Nguyen, T. N., J. B. Hubbard, and G. B. McFadden, J. Coatings Tech., 1991, 63(794), pp. 43-52 11. Rolfe, S. T.; Barsom, J. M. Fracture and Fatigue Control in Structures; Op. Cit; Chapter 8, p. 232 12. Szklarska-Smialowska, Z. Pitting Corrosion ofMetals; National Association of Corrosion Engineers, Houston TX; 1986; p. 3 13. Dickie, R. A. Prog, in Org. Coatings; 1994; 25; pp. 3-22 14. Reed-Hill, R. E. Physical Metallurgy Principles; 2nd edition; D. Van Nostrand Company, NY; 1973; p. 761 15. Pommersheim, J. M;. Nguyen, T.; Zhang, Z.; Hubbard, J. P. Prog, in Org. Coatings; 1994; 25; pp. 23-41 16. Tait, W. S. Corrosion; 1994; 50(5); pp. 373-377 17. Nelson, W. Accelerated Testing, John Wiley and Son, NY, 1990, Chapter 3, p. 129 18. Spray Technology & Marketing, June 1997; p. 10 19. Tait, W. S. in Corrosion Control by Coatings; ACS symposium series, American Chemical Society, Washington, DC; (in press); Chapter 18 20. Tait, W. S. Spray Technology & Marketing, September 1997 21. Kapur, K.C.;Lamberson, L. R. Reliability in Engineering Design; John Wiley and Sons, NY; 1977; Chapter 11, pp. 314- 323

Bauer and Martin; Service Life Prediction of Organic Coatings ACS Symposium Series; American Chemical Society: Washington, DC, 1999.